About this talk
This talk presents the concept of a sovereign AI platform, emphasizing the need for control over data and compliance with European regulations, particularly in light of GDPR and the AI Act. The speaker highlights the importance of European high-performance computing initiatives and introducing an AI infrastructure that includes federated environments and multi-tenant capabilities. They discuss technologies such as Slurm for job scheduling, the integration of container runtimes, and the significance of building scalable AI platforms that use both denser computing resources and energy-efficient architectures. Additionally, the speaker explains how the architecture can support advanced use cases through federated queries, data contracts, and the iteration of AI models using proprietary datasets. Finally, they touch on the long-term sustainability of AI and its alignment with green technology.
Full transcript
[music] >> Welcome to JCON. >> [applause] >> 36 years ago, I was standing in front of an audience like this. And I showed him what you can do with a computer that's 120 km away. It was the fastest computer on the planet called Cray 2. Who remembers that one? All right. And sorry, I'm not What? Interference. I don't want to pick on you, but there was a
professor >> [snorts] >> I worked at the university. Professor stood up and said, "Mr. Strauss, nobody needs that. I can send a floppy disk." That was the birth of the internet. Then the next 20 years we saw these waves, these ups and downs. Now, 36 years later or over 30 years later, we're seeing similar waves. They're called AI. A lot of excitement. A lot of yeah. And
a lot of announcements and people go like, "I am AI." Whoa, hold on a minute. So, question number one, how many of you are developers? Good. How many of you are developing for their own company? So, you are the end user. All right, that's about 1/4. How many of you are developing code for somebody else or for your customers? All right, so 25% 75%. Just wanted to
know. Um the important thing I want to talk about in the next 20 minutes is how can we build a sovereign AI platform? Why is this important? Page 185 in four-point letters on any cloud service provider agreement says "Well, we're obliged to give your data to the agency of if they want." The European Union decided, "No, that's not what we want. We want to give people in
Europe the ability to be sovereign." I'm going to come back to what sovereign means because you always need to define the context before you speak any words. Cuz everybody can say, "I'm sovereign." So, Andreas, what's sovereign for you might be different what's sovereign for Mark. Mark Mark Mark Miro Miro So, this is why it's so important. This is the definition that we have from the European Commission
their plans to build out a European AI cloud. Who heard about the European high-performance computing joint undertaking? Who knows them? EuroHPC JU? Okay, couple of you, good. Yeah. So should say hello from Anders. our goal is to do this. Why is this guy from Dell now talking to you about European Union and and what their goals are? I am responsible to build out the European ISV ecosystem
for our company, for AI. So, finding the people that are European-based under European jurisdiction to to bring software and eventually hardware to you or to your companies. what you see here is control over your stack is very You need to own your stack. Sovereign computing cloud infrastructure sovereign models and data spaces. All these various projects that we all hear about and whatever, they all play a role
in this, yeah? And governance and compliance, you're not only we've got GDPR rules, we also got the European AI Act and I'm very proud of that. There was a lot of people in the last couple of years said, "Yeah, Europe is falling behind. This AI Act is stopping us doing the the the good things." No, it is the right thing to do. Why? The base foundation of
all of this is if you run an AI use case you're responsible for it. Responsibility, ladies and gentlemen, is because you cannot relay the responsibility and deflect it to, "Oh, no, that was not my fault. It is OpenAI's fault." Your company, your CEO is personally liable, period. Done. Responsibility in good old terms. Good. So, oh thank you. >> [laughter] >> Why is this on in presentation mode?
Come on. Teams. All right. I skipped one slide. There it says a couple of European projects. You're going to get the slides, you can read that all in the news. Very important, this is the three major aspects that I as sovereign. control it. You control your stack, your software, your data, and your platform, period. It's not Sam Altman, it's not somebody else, it's you. If you have
control, you can run it. You own it. It's not somebody else who owns it. And it's on your terms and conditions of your company, of your executives or of you. Okay? when we build a platform, I look at a camera. I love it. It's like >> when you're building a platform you're techies, so you probably understand what this is. Yeah, a lot of people will talk about,
"Oh, yeah, I got I got a AI platform for you. I got something." This is what we're building. European Commission has said, "Oh, we're going to build this federated environment." Do you all understand what federation means? Of 13 AI factories. I count them in the in the power usage, about 10 to 20 MW. Five giga factories, that's 200 MW plus as a data center. And then for
every factory, you get about four to five antenna factories, smaller ones, 1 to 2 MW. And they all will be connected. So, you can schedule a job in one. If they haven't got the resources, they can relay that job into another one. 50% of that capacity will be coming to the public. So, your companies, the European citizens. 50% will stay academic. Different to what HPC was in
the past where everything was an academic supercomputer. So, that needs multi-tenancy, cloud service provider technology, whatever. This is the foundation of it. This is how it looks like. We're going to build an infrastructure layer compute, storage, networking with all the security. By the way, AI so funny. Who heard about McKinsey and Deloitte and them them losing the data? 1 2 3 security hacks that we know for
25 years. Yeah? How can it be that somebody launches an AI chatbot puts it online and then somebody can do SQL injection and copy the whole database of of their Remember, you need security. So, the good old, yeah? You're not a coder, you're a developer. You need to think your architecture. You need to think [clears throat] what's right. You need to think what's good. So going back
security is part of it. You need the infrastructure segmentation layer compute, storage, network security, load balancing. You need to manage it because it's a multi-tenant environment. And multi-tenancy, ladies and gentlemen it has a definition. It needs to be a secured environment because you might have a defense customer sitting next to a finance customer on the same platform. So, it's not like, "Yeah, I can separate namespaces in
in kubernetes. It is real Then on top of this, we built these segments and an operating layer this is combination of bare metal HPC like services where you can run Slurm jobs. Anybody who knows Slurm here? Right, two. Slurm is the mechanism how you you run jobs in a big supercomputer. It's batch Good old you know, it's like very very old style batch jobs. You have a
job, you put it in, you log into a node via SSH you it's like and you run the job, job's finished, you copy the data, then you you process the data outside later. That's basically the fundamentals of Slurm. And if five, 10, 100 people are using the same supercomputer, you need to make sure the scheduler uses all the resources in the right way. That's Slurm. So, you
got a mix of Linux operating system bare metal hosts that people will bring their own operating system all the way to container runtime environments. On top of that, you need a proper job scheduler that can use all these resources and this is important for the federation cuz you might not have the capacity here. It might be that this part of the job the return of SOA, who
knows SOA? >> This part of the job runs somewhere else. But it's still one application. Yeah. So, that's what you're doing in the job scheduler and then you built automation platforms and the API layer so that somebody can come in and use your large language model or small language model as a service and run a chatbot or a rack reality augmented generative AI. So, feed your own
documents into a model. So, that is basically how this thing looks like. So, it's an architecture that we all should know. And it's not a single vendor, single company, I'm going to build it for you all because I'm in video. Sorry. Didn't say that. So, when we look at the main aspects of an AI platform, we need to think about three things. The AI infrastructure. And since
you are software developer the word infrastructure for you is different than to an IT person. You think in infrastructure as libraries, compilers, tools, code, whatever. When an IT person thinks they go server storage networking, GPUs, CPUs, load balancers. imagine what happens in AI when these two talk to each other, they they don't understand each other. So, it's very important that we understand what AI infrastructure means. AI
data you're developers, you know it. All these large language models have been trained on the same data sets in the internet. There's no private data sets they have been using for training them. So, they're all pretty much the same. Yeah, they've been trained on more documents and more you know you know, if you four billion parameters or 400 billion parameters is like yeah, you you fed it
with more documents, but it's still the same fundamental data. But the one thing that is very important, this is what a lot of people forget is if you want to have a real intelligent or it's not really intelligent. It's doing statistics. You need to train it with your own data, your company data. And this is why I don't want my corporate IP to go out and train
some of them's models. So that the rest of the world knows all my IP. Cuz that's what's happening. Even if they say, "No, no, no, we're not training it. You're just inferencing." Inferencing is when you use the model in a chatbot or you in an application. But still every piece of data you you you shoot into it is being logged somewhere and will be reused. So, it
learns while it's talking to you. Call it the reinforcement learning for all of you in in in the AI space. And that's like, "Nope, that's why my data needs to stay where I where I want it." And location, very important. Jurisdiction is one thing, regulatory environments, and the other thing is where is the processing of my AI and my data being done? Do I have my own
capacity? Do I have a data center that can do that? Think about when this largest supercomputer I talked about it was about that size. You all know the picture, you could stand inside. I have that picture. And it was about it consumed about 200 kilowatts. We're back in that world, a single rack of computers consumes 180 kilowatts today. Most of the data centers can do 25 to
30 kilowatts in a rack. So, imagine you take a hall of this which is full today with servers you typically have, you put all these new boxes in it's going to be an empty hall. And this single rack is liquid cooled, it weighs 1.3 tons. So, if you got the chance to see a data center where somebody has put in these new liquid cooled systems, you see
a a lot of empty space a lot of metal plates around it because that weight on a raised floor is quite something. Yeah. Again, that's what you need to consider. What do use case builders need? These people like you you use specific tools and and Most of the use case builders and I interviewed about 509 of them in the last eight years they use things like Amazon
Bedrock or Sagemaker. They use these these environments to pilots and and minimal viable products. And all of them said the same, "Johan, my problem is that I'm now talking to a customer and and show him oh, this is this is your this is your AI, this is yours." And the customer goes like, "No, you're not going live with that. I'm not allowed to put that on Amazon
by because I'm a bank or whatever. You know, I got a regulatory environment or my company does not allow me to do that because it's corporate intellectual property. So, what we need to build in the enterprise is that platform so people can develop here and then say, "When I go live, I bring it here." So, let's look at the fundamental building blocks of that. I'm very proud
that in the European Union we've we've doing this for quite a time. I wrote about AI 35 years ago first time in my life. Functional blocks of an AI platform. on the left side, you got your data. Data can be in any source in any form. Whatever you see out there. That needs to come into the first functional block called a data platform. And this is what
Mark is going to talk about. How much time do we have? Two more minutes? I need five. data will be coming into that data Major functionalities of that data platform and this is very important that there's only three major things you need to consider. One is you need data contracts. Data contract means it's not a product, it's it's like what am I allowed to read from you?
Can I copy your that data? Can I copy your spreadsheet? Am I Are you masking something out which is personal data or whatever? What am I allowed? That's my data contract with you. Second is federated queries. I haven't got a running microphones so nobody can answer. Any idea what federated queries mean? No? You leave the data where it is and you remotely query it. Think about it,
we're in an exabyte world. The example of Dell. On the left side are sources of data in the company sitting in 48 different locations in our data centers and in cloud providers and in our partners is about one exabyte. First of all, try to copy a petabyte of data from A to B. Good luck. So, data and location is important. That's why federated queries are important. You
query the data. To give an idea, the exabyte here turns into about 20 petabyte on the data Cuz I only need that data that I need for my AI project. Then I embed that into vector and graph databases so that AI can process it and a model can process it. Then you need a a model hub. You all know examples of it, GitLab, GitHub and co. Because
you're experimenting. Doing AI projects not like writing software where you do you do your structure and you code. You do experiments. Cuz you got five different with the same data set and then you start experimenting which one gives you the best outcome. So, you need to be able to do that, to customize it, version, to do governance, risk, and compliance. You need to In the European Union,
we need to show this is the data I'm using, this is the model I'm using, this is the process I'm using, and so on. This needs to be documented as part of the AI. Then you need that data science tools. You're all aware of them. And that's the free functional box that the data science people need. And then you need the virtual infrastructure, compute storage, network, and
totally software defined. Why? This is not SAP HANA or something. The next model has complete different resource requirements. Either on storage, on network, on CPU, on GPU, it has different requirements. So, you need to be able to to do that. And then you need to run it, and that's the operational part. That's the five functional blocks. When we're building the five functional blocks, there's a couple of
subcomponents we need to build, but the very important thing is what you see on the right. A, we want to be silicon independent, and the European Union was very clear, we also want to include European technology. RISC-V, anybody knows about this? Okay, one. RISC-V is one of the technologies to build European CPUs. Really? European Union? Yeah? Are they available in big? No. So, I think about here's
a thousand racks, and we got two racks with these RISC-V processors as part of these big platforms. But that's like you're you're independent, if you're Intel, AMD, RISC-V, whatever, this Kubernetes flavor on top needs to manage it all, the multi-tenancy platform needs to manage it all, and then you get this So, when we're looking at these building blocks again, simplify the architecture, you got server storage networking
and co, all the way up to the container environment. That's what we're building as Dell. We're not just building hardware. Then we bring ecosystem partners in to build the higher layers. And you see some examples here, and this is my job. You're the the people that do these AI platform here. Yeah? Here's a couple of examples. Scirocco is one of them. Deep Set, Germany. Another one. Confidential
Mind, Finland. That's my ecosystem that I'm bringing into into the company, Dell, and over time, you know, when we done a lot of projects together, we're going global. First is we're doing projects together, then we start to integrate this, then we start to test everything, and then we go and global. Last thing to discuss. So, when we're building that platform, so we have the infrastructure layer, we
got a combination of Slurm HPC clusters. Why do you need HPC? AI works best if you have simulation at as well. So, you want to simulate digital twins and everything. That's high performance computing. This is why sometimes you need Slurm in this this case. Then you different Kubernetes cluster by design, not a single one, because you need to be secure. You need a control plane on top,
and then you put everything else into the tenant space. There was a lot of confusion. For example, any Red Hatters here? OpenShift AI and OpenShift. Even Red Hat people got confused with this. OpenShift is the Kubernetes flavor. OpenShift AI is what you see on the left on tenant A. That's the AI platform layer on top. Sets in the tenant space. Very important. thank you very much for
listening. That is the start of this journey. technology, European software is what I'm standing for until I go to retirement in five year four years from now, sorry, four years. Um but as I know for at four times the speed of the internet, I'm going to make it. It's going to be there, up and running. Marcus, you're the next one. Over to you. Thank you very much.
I'm going to be around for another couple of hours. >> Super. Thank you. Great. Thank you on um awesome awesome stuff. And now we combine all of this with Java, and this is what we are talking about. So, we are partnering with Dell and include or we bring a whole Java stack on the table is a whole AI platform software platform and unifying everything you need for
building AI applications in pure Java. It integrates with Java, and uh it should be great for the Java community because of it's it feels like the part of Java. And now I'd like to introduce you how the system works. an AI knowledge fabric, so it delivers the whole infrastructure you need for building these applications in pure Java, because everybody is talking about Python when you talk about
AI, and so we're discussing can we do this better in in Java? With our great language, the great platform, the great ecosystem. we built we built such a platform. It it brings two USBs on the table. Um high efficiency, it saves you up to 80% infrastructure cost. Um or let's sell it for Dell, you get five times more AI power for the same price. And we can
process um data volume in in petabyte size. scalable. And we need this for AI brain. This is kind of like kind of like AI brain, and here we build the AI long-term memory. We're talking about long-term memory. Um why did we build this, and when did we start? So, some information about the history. So, we started this project 10 years ago. We worked for 10 years on
this project. Uh of course, we didn't know that AI will be a topic in 10 years. But what we were doing or we were starting this project um working with Java object graphs. So, storing Java object graphs, loading Java and managing state highly efficiently in Java. So, this was the the project all about. What what is an object graph? So, highly connected objects can grow infinitely, um
can be queried easily, um super efficient by using plain Java. So, we built three Eclipse projects, which are which are the core of this technology. Eclipse Serializer. everything started with this serializer. Why build we a new serializer? So, Eclipse Serializer is highly secure. You cannot inject malicious code and execute malicious code through deserialization. So, it's highly secure. If you're looking for high security solution in Java, check
this out. It's completely open source. So, we can serialize highly complex object graphs efficiently. Then we build a storage engine on top of A storage engine that allows you to store any complex object graph, even with circular references, you can store it persistently on disk, re- reloaded and and restore it in memory at any time, completely or just partially loading subgraphs. We are using lazy loading highly
efficiently. Then the cloud came up, and so we were talking about distributed applications. Then we built on top of that Eclipse Data Grid, an infrastructure as code that generates the whole infrastructure to distribute your Eclipse Store So, it distributes your Java object graph through multiple JVMs, hundred, even thousand JVMs, so we can distribute your object graph infinitely. You can query it by using Java Streams API. Um
and now the last month we were spending a lot of time to integrate JVector, which is a highly efficient vector search written in Java. And now you have Eclipse Store, um we you can use it as an embedded in-memory vector database. Uh it allows you to build all of this in plain Java. And with Eclipse Data Grid, you can distribute this. So, now some years ago we
were attending our first AWS, sorry, our first AWS uh training course with Adam Bean. So, and we So, it it bring it bring us to a a new concept of processing data. It was planned as a next generation Eclipse Store or next generation Eclipse Data Grid. even more efficient, even even more powerful, and we recognized this fits great to AI and solves the problem that AI has
today. And now this software is available. By the way, all the Eclipse projects, of course, are under EPL. All features are included. So, check this out. It's completely free and allows you to build commercial software. We have, of course, huge problems with energy consumption with AI. Everybody knows this. Okay. now the goal is to build a high-efficiency AI platform and not AI in terms of LLM or,
yeah, the intelligence of AI. It's more like how we store data long-term. how we build um an AI long-term memory. And here we saw that the object graph model is perfect the perfect knowledge is data that is connected, highly connected, like a like a human brain. So, like a graph, like a knowledge graph, like you have in a graph Um when we store data long-term, we today
use a vector database, of course. we also can use caches for short-term memory. And now imagine we can now unify all of these solutions in one plain Java native Java native solution that allows you to use a a knowledge graph that grows infinitely. You can query it by using vector search, graph traversal, full-text search. Everything combined in pure Java. So, why can't we do that with traditional
databases today? We can do this. But, they are very expensive. Why? Imagine you have an infinitely growing knowledge. We want to process this data by using a static database. So, databases are still monolithic, static. And they have problem. Um it's a coupling between the data you want to store, the memory you need, and the processors, the CPUs you get. So, the more data you want to process,
the more memory you need, the more memory you have, the more CPUs you get. Even if you don't need the computing power. So, this is the reason why monolithic systems waste up to 80% computing time by doing nothing because you don't need it at the moment. So, it's not energy efficient. And such systems are very expensive because memory is extremely expensive. So, price tripled and it's skyrocketing.
And these systems become critical beyond 100 TB. But, when we talk about AI knowledge, how should we process data in in petabyte size, petabyte volume? So, it becomes, yeah, extremely expensive. it's impossible. How can we process data uh petabyte data sizes today? So, today we store this amount of data in in cheap data storage like uh data warehouses or S3, whatever, and store it on disk, where
it's very cheap. And then we extract some some data, some important hot data, push it into a vector database or a graph database, which is very expensive. And then maybe we can process 10% or 5% of the whole Um and the all the other all the rest is stored in an archive and we cannot touch it. If you want to touch this data, it's very expensive because
then we have to extract this data. Takes sometimes hours or maybe a day to extract this data. And we need another cluster and so on. And such machines, yeah, cost up to 1 million or even more. So, it's it's mostly too expensive uh for mid companies and even for the big And the hardware is not available today. Um so, how we solve this problem? So, this problem
can be solved by just using a completely new architecture. And uh this architecture was not invented by us. Uh it was invented by AWS. So, the the functional principle is simple. Uh it's called serverless functions. Instead of a huge server machine, we use tiny functions, super tiny functions that they run only when we need the function, and we can use it as a data storage function and
it fetched data from disk, returns to the app, and shut down. And consumes CPU power only when it's And we don't need it, if we don't touch it, shut down, does nothing. No CPU, no cost. Great. Wrong direction, sorry. So, with serverless functions from the functional principle, we can navigate through an endless growing AI knowledge, highly connected data by in by invoking uh tiny services that delivers
data, shut down, and this is highly efficient. With this functional principle, AWS showed that we can save up to 80% infrastructure costs by design, but by by using this architecture. Who of you uses microservices and serverless functions today, already today? Okay. Yeah. serverless functions have some drawbacks uh because they are quite slow. So, startup time is not super fast. when when we when we so, cold startup
Sometimes seconds. And imagine if you want to get access to a to your AI knowledge, takes seconds, this super slow. We want to get access to a long-term memory in milliseconds, like the human brain. I- if everything we talk about right now is short-term memory. This is GPU. This is LLM. If I ask you, how was your first day at school? You have to picture in milliseconds,
right? How does it work? If we use a data lake, your IT division would say, I can give you the answer in 4 hours. With serverless functions, we can give you the answer in 3 seconds. Better, but not great. Now, we build a system that gives you the the information in milliseconds. How does it work? We use this functional of serverless functions. We call it storage cell.
we build a Java application that groups all of these cells, or not all, let's say 100,000, sometimes 10,000 in one Java application, and run this group on Kubernetes. But, Java is so slow and so legacy. It's not suited for AI. We convinced all of our investors because we were able to show that Java is super fast. With GraalVM native image, we were able to compile this this
cell group and start the cell group on Kubernetes, and it consumes less memory. So, it's highly efficient. And it's a great use case for GraalVM native images. So, if you if you need um Java application that starts in instantly, use GraalVM native image. Awesome. Now, we can combine a infinitely growing knowledge graph with this architecture. So, imagine we have infinitely growing knowledge, we traverse through this streams
API, graph API, uh or or vector indexes. Everything combined in one Java statement. If you touch an object, there is a function invoked, it's started in milliseconds, fetches the data from disk, returns to the AI app, and shut If your AI knowledge does nothing, like a human brain. So, a lot of people do Don't think anything. Your brain does not consume any power, any any energy, it's
super energy efficient. Same here. Your server will do nothing. You can spend the CPU power for other processes. want to traverse through a highly complex graph, a lot of functions are invoked, do the job, and shut down, and do nothing. So, with this infrastructure, we can save 80% or even more infrastructure costs. Um and it combines everything in pure Java. So, Uh, some weeks ago, Google announced
a cool technology where they can uh reduce the the context window uh with Turbo Quant. Uh, it it's a breakthrough. And uh this is a cool technology uh breakthrough in the short-term memory. And now with Scirocco we deliver the counterpart for the long-term memory. And now we deliver all of this uh to the Dell AI infrastructure stack provide the high efficiency server machine for your data centers
for your AI application uh that can be run in your organization as Yuan explained. And this is what we deliver now with If you're interested, check this out. So, the whole Eclipse Store the Eclipse Store Eclipse Data Grid Stack is included. It's open source. You can also use this as a as a as a cache as an AI cache as a long-term memory. Everything included, everything in
Java. And today we have a another session at 2:30 where we show some code and some examples uh how how this works, how we can query such a knowledge graph. Uh, and Florian will show some some All right, how much time do we have? Oh. Nice. So, uh what we are working on right now is um we work on deterministic um auditable And uh the next step
the next evolution step of this technology is uh autonomous AI knowledge. Imagine there is a huge object graph a a huge knowledge graph infinitely growing and there is a process running behind the scenes connecting your AI data your long-term data automatically. And it's reorganizes everything automatically constantly like a garbage collector behind the scenes for AI knowledge. Uh this is what we're uh working on um in the
next upcoming month to build And now with with Dell we want to ship this to the industry and uh making the global center for sustainability for sustainable AI. And everything built on Java. And the cool thing is you have not to new query languages new data structure. You can use your core Java knowledge your core Java Java ecosystem. So, um all the the Java libraries frameworks can
be used and uh in one solution. If you have any questions you can find my team uh on the the Scirocco booth. for people who know uh already Eclipse Store MicroStream uh this is a further development of a MicroStream and Eclipse Store um and uh we have uh announced the version uh we have published the version four some weeks ago. So, it's it's further developed uh and
with with more and more progress. Okay, if you have any questions? No questions? What will be the main goal initially to use technology to replace? What would you replace first and foremost if I were to use the technology today in my enterprise? Basically you can integrate this in your existing infrastructure. You can choose the parts of the stack you need. Um if if you want to build
your AI application completely with Java um you don't need an external vector data database anymore. You don't need an another graph database anymore. You can do this completely in Java. And so it unifies everything you you need in in core Java. Saves you a lot of infrastructure costs, license fees uh and of course uh costs for maintenance. So, the idea is to simplify the infrastructure as well
as providing technical benefits for the application. Exactly. All right, I just want to take over for 30 seconds. This this is very important for all of us. The job of a CTO is explained to everyone even if you know it you know to so that we all on the same page. The short-term memory called KV cache is when when you do your your your chat with your
Anthropic or whatever you know that's that's that's the cache. But that only stores your actual session. The graph knowledge that knowledge graph is when you take all that data that you produced and you did your vector search and then you build the knowledge. So, an output of a real chatbot of your own data will need that knowledge graph and that's the long-term And when you build this
in classic I can give you the example. If you're doing this on a PostgreSQL with vector extension in Apache for five petabytes I know it because we've done it internally. That needs about 200 servers with high memory. So, the cost he was is real. Sustainability in AI you we cannot argue that GPUs you know they are at 750 watts a day they are going to 2 kilowatts
in in about 2 years. These these beasts need a lot of power. Unfortunately optical processors also I wrote about that 30 years ago still not existing. Um once they are there there there might be a reduction. Until then we got transistors and they need a lot of power. So, the last thing you want to do is have this GPUs running in your data center consuming hundreds of
kilowatts and waiting for data. This is why it's important you know the milliseconds and the nanoseconds when you're bringing that not graph knowledge back into the large language model is important. Otherwise it's cycling and it's it's producing nothing than heat. Yeah? The the green aspect of AI is coming is like the the the faster you feed the data in the more it can produce. We can't negotiate
that you know I I used 20% of the power in the GPU then it will run on nothing. Yeah? But making it fast so that it works more is the important thing until we got new technology which hopefully will come out in the next 10 years. There's one more question. I have a question. I will repeat the question. Okay, how do you maintain the data consistency? Because
the industry needs exactly the consistent data. In case that you apply the LLM this data all this data are probabilistic. Because you Mhm. There's the border and this is not pro possible to recover anymore. So, you need to maintain the consistent data. Very very good question. Let me try to rephrase the question. Have we got still a couple of more minutes? One more minute. So, what one
of the important aspects that Miro Miro Yeah. Remember? Yeah. Important aspects of of AI models is everything is is statistical. Yeah? So, if you if you feed a model with data today the output now is this. You do it again in 5 hours the output is different. Reason for that is because it's just doing statistical multiplications and calculations. The knowledge graphs the long-term memory helps you to
be consistent in that data. Cuz you're not just doing statistical calculations. But again the data problem is the biggest problem we're having in this industry. we can't change the models. They're doing statistics. They're not intelligent. Sorry. Also we call it AI. They're not intelligent. They're doing So, taking the audibility that Marcus was talking about into your vector and graph databases is super important. And a lot of
technology that is out there that is being used today does not do that. So, you cannot reproduce the same flow with most of the systems that are out there in place today. the answer to the to the question with your workshop with a bunch of us. But this is like this is one step forward in making things auditable and reproducible. You know, some people the marketing people
then call it explainable AI if you heard that that thing. All right, um I think we're out of time now. Thanks a lot. Thank you. Keep Keep asking questions. Find Find us on LinkedIn.
